Journal of Power Sources 196 (2011) 208–217

Contents lists available at ScienceDirect

Journal of Power Sources

journal homepage: www.elsevier.com/locate/jpowsour

Modeling and control of tubular solid-oxide fuel cell systems: II. Nonlinear model
reduction and model predictive control
Borhan M. Sanandaji, Tyrone L. Vincent ∗ , Andrew M. Colclasure, Robert J. Kee
Engineering Division, Colorado School of Mines, 1600 Illinois St., Golden, CO 80401, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper describes a systematic method for developing model-based controllers for solid-oxide fuel
Received 13 May 2010 cell (SOFC) systems. To enhance the system efficiency and to avoid possible damages, the system must
Received in revised form 18 June 2010 be controlled within specific operating conditions, while satisfying a load requirement. Model predictive
Accepted 21 June 2010
control (MPC) is a natural choice for control implementation. However, to implement MPC, a low-order
Available online 3 July 2010
model is needed that captures the dominant dynamic behavior over the operating range. A linear param-
eter varying (LPV) model structure is developed and applied to obtain a control-oriented dynamic model
Keywords:
of the SOFC stack. This approach effectively reduces a detailed physical model to a form that is compat-
Tubular SOFC
Physical modeling
ible with MPC. The LPV structure includes nonlinear scheduling functions that blend the dynamics of
Model predictive control locally linear models to represent nonlinear dynamic behavior over large operating ranges. Alternative
Nonlinear system identification scheduling variables are evaluated, with cell current being shown to be an appropriate choice. Using the
Linear parameter varying models reduced-order model, an MPC controller is designed that can respond to the load requirement over a
wide range of operation changes while maintaining input–output variables within specified constraints.
To validate the approach, the LPV-based MPC controller is applied to the high-order physical model.
© 2010 Elsevier B.V. All rights reserved.

1. Introduction tail-gas combustor, heat exchangers, power conditioning unit, etc.

[3].
This paper is the second of a two-part paper that develops a In a system such as the one illustrated in Fig. 1, vaporized fuel
model-based approach for control of a tubular solid-oxide fuel and air are converted within a CPOX reactor to produce a syngas
cell (SOFC) system. The companion paper develops the physical mixture (i.e., a mixture of H2 , H2 O, CO, CO2 , and N2 ). The syngas
model and applies linear model-reduction techniques near par- enters the anode side of the SOFC stack. Preheated air enters the
ticular steady-state operating points (OPs) [1]. The present paper cathode side of the SOFC stack. Air and unspent fuel leaving the
continues to use the same physical model, but develops linear stack are mixed and burned in a catalytic tail-gas combustor. Hot
parameter varying (LPV) methods to extend the reduced models gases leaving the tail-gas burner are used in a recuperating heat
over large ranges of operating conditions. Finally, the reduced mod- exchanger to preheat the air entering the SOFC stack. Raw power
els are used to develop and demonstrate model predictive control from the SOFC is processed through power electronics before being
(MPC). delivered to the application load.

1.1. SOFC systems 1.2. Why is a control system necessary?

Fuel cells are devices that enable the direct conversion of chem- To operate at multiple power levels, fuel cells require a control
ical energy into electrical energy, with a theoretical conversion system to balance the fuel and air supply, as well as the electri-
efficiency that can be much higher than for heat engines [2]. As cal load. This is important both to achieve high efficiency, as well
illustrated in Fig. 1, the systems are an integrated combination as to avoid operating conditions that can damage the fuel cell, such
of contributing components including the fuel cell stack, together as excessive temperature or temperature gradients, catalyst coking,
with balance of plant (BOP) components that include air blowers, and anode reoxidation. Especially for small portable power applica-
fuel pump, fuel reformer (e.g., catalytic partial oxidation (CPOX)), tions, the SOFC must deliver power profiles that meet the transient
load demands. Consider, for example, an auxiliary power unit (APU)
that is designed to satisfy the hotel loads for the sleeper cab on a
∗ Corresponding author. Tel.: +1 303 273 3641; fax: +1 303 273 3602. long-haul truck. Depending upon the activities and appliances in
E-mail address: tvincent@mines.edu (T.L. Vincent). the cab, the power demanded from the APU can vary considerably

0378-7753/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.jpowsour.2010.06.075
 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217 209

and identification of an SOFC stack. Arriagada et al. [9] develop an

SOFC model based on an artificial neural network (ANN). Huo et
al. [10] develop a Hammerstein model consisting of a static non-
linear block followed in series by a dynamic linear block. Wang et
al. [11] consider a data-driven modeling approach for identifying
SOFC systems.
All the foregoing examples have limitations. In some cases, even
when nonlinear model structures were used, only static (steady-
state) models were considered. In other cases, transient models
are used, but they are only fit near a single OP. Such modeling
approaches are generally less useful for control because changes
in the required load on the SOFC stack cause both the steady-state
equilibrium and dynamic response to vary significantly at different
OPs.
Recent research by Hasikos et al. [12] for PEMFCs considers mul-
tiple OPs. They generate a steady-state database of manipulated and
controlled variables and then fit an ANN model to that database.
Fig. 1. Schematic of an SOFC system that is typical for mobile applications such as After static modeling, dynamic behavior is added to the model by
an APU. including a finite impulse response (FIR) model of the system at one
of the OPs. Chen et al. [13] present a multiple-model approach for
a PEMFC/ultracapacitor system. They use fuzzy clustering to char-
over time. Although such systems are usually designed with bat- acterize and identify multiple linear models of the system, which
tery storage and power electronics that seek to limit the transients are selected based on operating conditions.
required from the SOFC, the fuel-cell system must still respond to a
range of transients. Moreover, if a control system can enable the fuel 1.4. Prior control research
cell to respond to load changes more quickly, storage requirements
are reduced, decreasing overall system cost. A variety of control approaches has been proposed for fuel cell
Any control system depends upon actuation and sensing. SOFC systems, ranging from simple linear state feedback controllers to
actuation is usually accomplished through a combination of cell more advanced control techniques such as predictive controllers.
operating voltage, fuel flow rate, and air flow rate. Sensing is usu- Li et al. [14] used a state feedback exact linearization approach.
ally accomplished with thermocouples that measure temperature. Pukrushpan et al. [7] considered dynamic feedforward and state
Pressure and flow-rate measurements are valuable, if available. feedback controllers along with the linear quadratic (LQ) method-
Measuring chemical composition is especially valuable but usually ology.
difficult and expensive. Thus, chemical information usually must be Several authors, including, Danzer et al. [15], Yang et al. [16],
inferred from other less-expensive measurements. There are usu- Jurado [17], and Huo et al. [10] considered using MPC for fuel-cell
ally practical constraints on actuation and on system performance systems. MPC uses a system model and an online optimization
to avoid damage to the cell. For example, the operating voltage may to calculate appropriate actuation commands. This is an attrac-
be constrained to remain above 0.6 V. The cell temperature may be tive option for SOFC systems, because MPC can directly handle
constrained to remain within a specified range. Fuel utilization is constraints on both the input and output variables. However, in
usually constrained such that some unspent fuel remains in the most prior implementations of MPC, control is only demonstrated
anode exhaust. A sophisticated control system is needed to meet around one particular OP.
constraints on actuation and response while guiding the system Hasikos et al. [12] developed an integrated optimization and
through power-demand transients [4]. control tool for PEMFC systems. After fitting a radial basis func-
tion (RBF) neural network model to a steady-state database, an
1.3. Prior modeling research MPC methodology was designed based on a FIR model of the sys-
tem at one specific OP. This results in relatively poor transient
For a model to be useful for control purposes, it must be of response with long rise time and/or output jitter when the system
low complexity and yet capture the dominant dynamic behavior is operated away from the nominal OP selected for FIR modeling.
of the system. Control-oriented modeling ranges from first princi- Chen et al. [13] considered a multiple-model predictive controller
ples physics-based component modeling using flow characteristics, for a hybrid PEMFC system. In multiple-model MPC design, an
mass- and energy-balances, simple diffusion and heat equations to upper-layer adaptive switch is added that determines which of the
data-driven methods that fit low-order empirical models to exper- models should be used within each sampling period. However, the
imental data. implementation can be computationally expensive and switching
Examples of physics-based low-order modeling include, Hall between linearized models can cause unintended perturbations.
and Colclaser [5], who develop a transient model for a tubular
SOFC, Padulles et al. [6], who propose an integrated SOFC plant 1.5. The present approach
dynamic for power systems simulation and Pukrushpan et al. [7],
who develop physics-based dynamic models of proton exchange The present approach to control of the SOFC stack is to develop
membrane fuel cell (PEMFC). an MPC controller based upon reducing a complex physical model
One difficulty of physics-based low-order modeling is that when of the stack [1], with the low-complexity model being specifically
choosing which physical effects to include, it is often difficult to tailored for real-time optimization. The model reduction is based
balance between computational simplicity and physical fidelity. As on a specific LPV model structure that continuously blends multi-
an alternative, many investigators adopt a data-based approach, ple linear models according to a so-called scheduling parameter.
which begins with data from either a high-order model or an exper- A velocity implementation is used to ensure that only small-signal
iment and fits a low-order empirical model. For example, Yang information is applied to the linear models, thus assuring bumpless
et al. [8] consider a modified Takagi-Sugeno fuzzy (TSF) modeling transfer between OPs [18].
210 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217

Fig. 2. System block diagram showing input–output allocation of the control struc-
ture. Note that the tubes in the SOFC stack are in a parallel configuration in regard
to the fuel and air, but in a series configuration electrically, so that the stack voltage
is the sum of the tube voltages.

Fig. 3. Balance of plant subsystems.

Model reduction proceeds in two steps. In the first step, a set
of OPs is selected. At each OP the physical model is systematically
perturbed and the model response is recorded. A linear model can the order of one second or less. Characteristic times for thermal
be identified using the relationships between perturbations and response are much longer, on the order of minutes. Because of
responses. As discussed in the companion paper [1], the linear mod- relatively fast characteristic response times and strongly nonlin-
els capture the small-signal behavior of the SOFC stack near the ear behavior, controlling electrical and exhaust characteristics is
selected OPs. This linear identification is followed by a second step more challenging than controlling the thermal behavior. The MPC
in which nonlinear scheduling functions are identified that mod- approach is particularly well suited to controlling the cell voltage
ulate the outputs of the small-signal models in a manner similar and fuel flow rate. Because of the slower thermal characteris-
to gain scheduling. A particular contribution of the present paper tic times, a more traditional proportional-integral-derivative (PID)
is a novel non-parametric approach for identifying the schedul- controller is sufficient for temperature regulation.
ing functions. In the LPV scheme, one of the measurable output
variables is considered to be the scheduling variable. Alternative 2.1. BOP components
choices for scheduling parameters, and combinations of schedul-
ing parameters, are investigated. An LPV-based MPC controller is Although the uncontrolled dynamics of the BOP components
designed, implemented, and tested. Controller performance is eval- may be non-trival, because the BOP components are under servo
uated by controlling simulations based upon the detailed physical control, the relevant closed loop dynamics (that is, the dynamics
model while the MPC controller exploits the low-order fast model from a set-point to a physical variable, such as air flow) can be
of the stack in its online calculations. Several demand trajecto- simplified from the point of view of the stack controller. Specifically,
ries are evaluated, considering rapid transients in output power it is assumed that the closed loop dynamics can be approximated as
while satisfying constraints on cell voltage, fuel flow rate, and fuel linear dynamical systems with first-order lag dynamics and unity
utilization. DC gain. For example, the fuel-flow subsystem is modeled as:

qd = fuel q̇in + qin , (1)

2. Control strategy
where qd is the desired fuel flow rate,  fuel is the time constant
As illustrated in Fig. 2, the system-level control strategy is hier- associated with the fuel-flow subsystem, and qin is the fuel flow
archical, with MPC used at the highest level. The SOFC system is rate command that is supplied to the BOP. Analogous first-order
comprised of the stack, shown on the right, and the BOP compo- dynamics are considered for the air-flow subsystem. Because the
nents. Fig. 3 illustrates the interior structure of the BOP block, which dynamics of the cell-voltage subsystem are very fast, it can be rep-
regulates voltage, fuel flow rate, and air flow rate. Since the tubes resented simply as a unit gain. The BOP control is not discussed
(cells) in the stack are connected electrically in series, the regulated further in the present paper.
voltage is the sum of the individual tube voltages. Local servo con-
trollers are used to achieve desired set points in these variables. 3. Stack modeling
It is assumed that sensors are available to measure cell current,
hydrogen mole fraction in anode exhaust, and temperature of the This paper concentrates primarily on the SOFC stack dynamics,
cathode-exhaust air. which is more complicated than the BOP dynamics. The SOFC stack
As described in the companion paper [1], the physical model physical model is discussed in the companion paper [1]. The exam-
represents a widely disparate range of characteristic time scales. ples that follow in the present paper are based upon a particular
Electrical response is essentially instantaneous (i.e., because elec- tubular stack configuration. The tubes are 15 cm long with an out-
trochemical double-layer charging is neglected, a change in side diameter of 1 cm, which is typical of tubes that may be used
operating voltage causes an instantaneous change in current). in sub-kilowatt APUs. A table of model parameters that describe
Characteristic response times for fluid flow and diffusion are on physical and electrochemical characteristics may be found in the
 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217 211

Fig. 4. LPV model structure for low-order representation.

companion paper [1]. For purposes of illustrating the model, the

post-reformer SOFC inlet fuel composition is assumed to be 38%
H2 , 3% H2 O, 1% CH4 , 19% CO, 0.3% CO2 , and 38% N2 . This is the equi-
librium mixture that results from the partial oxidation of methane
and air at 800 ◦ C with a stoichiometric ratio to partial oxidation (i.e.,
oxygen to carbon ratio of 0.5), to which a small amount of steam
is added to bring the steam composition to 3%. The steam is added
to prevent carbon deposition in the fuel cell stack. Fuel enters the
tube at 800 ◦ C and atmospheric pressure. The cathode air, which
flows around the outsides of the tubes, enters the stack at 550 ◦ C.
Both ends of the tubes conduct heat to the manifolds at a fixed
temperature of 800 ◦ C.
A typical stack can consist of 50–100 tubes, usually connected
electrically in series, which is the case here. The modeling assumes
an ideal stack in which all tubes behave exactly alike. Thus, net
stack power and flow rates are determined from a single tube sim-
ply by multiplying the results of the single-tube model. The series
connection implies that the current through all tubes is the same,
while the stack voltage is the sum of the voltage across each tube. In Fig. 5. Input–output map from steady-state simulation. (a) Steady-state input map.
(b) Steady-state output map.
actual operation, the system deviates from this ideal stack behavior,
as the tubes are exposed to slightly different conditions. However,
the idealized model is still valuable for control, as it represents the
behavior of an “average” tube in the stack.
The process for obtaining the reduced-order model is twofold. with different input values are recorded. Fig. 5 illustrates an input
First, a set of OPs is selected and locally linear models at each OP map and the resulting output map. In these simulations, cell voltage
are identified to model small-signal behavior. The linear models ranges from 0.65 V to 0.85 V in steps of 0.01 V. At each cell voltage,
are then combined, with redundant modes being eliminated via fuel flow changes between 2.5 mg s−1 (velocity = 0.247 m s−1 ) and
linear model reduction. In the second step, nonlinear functions are 12 mg s−1 (velocity = 1.19 m s−1 ). The corresponding steady-state
identified that modulate the outputs of the small-signal models in a outputs define a region in which, roughly speaking, cell current
manner similar to gain scheduling. The nonlinear functions depend changes between 10 A (per tube) to 55 A (per tube) and H2 concen-
upon a scheduling parameter that is assumed to be measurable tration in exhaust varies from 0 to 35%.
(e.g., cell current). The practical operating space should be more restricted than
Fig. 4 illustrates the LPV model structure, where uk and yk are the potential operating space shown in Fig. 5. To achieve high effi-
the input and output at sampling interval k, and z is the forward ciency and to minimize cell damage, the cell voltage should be
shift operator. The locally linear systems, Gi (z), capture the small- limited as 0.67 ≤ Ecell ≤ 0.79. Because of a need for fuel into the
signal behavior at m different OPs. The system G(z) represents the tail-gas combustor for cathode-air preheating (Fig. 1) and to limit
combined model after eliminating any redundant modes shared deleterious reoxidation of the SOFC anode, fuel utilization should
between the m different linear models. The variable  is a schedul- be in the range of 70–90%. In the present simulations, this limita-
ing parameter that is a function of the (measured) system output, tion corresponds to maintaining hydrogen in the anode exhaust as
and fi () represent gains that are nonlinear functions of . Because 3% < XH2 < 11%. The combination of these considerations limit the
each of the linear models captures only small-signal behavior, the cell current (per tube) as 13 A < Icell < 35 A.
OP needs to be removed at the input, and added back at the output. The practical limitations in the steady-state output space are
A so-called velocity implementation accomplishes this by applying illustrated in Fig. 5b as a shaded rectangle. Six OPs are selected
a first difference (1 − z−1 ) of the input, and integrating (1 − z−1 )−1 (shown as large red dots) along the output boundaries to represent
the output. the operating space. The selected output space is mapped back to
the input space with corresponding markers in Fig. 5a. The selected
3.1. OP selection OPs along with their corresponding output values are also listed in
Table 1(a) and (b), respectively. In the following sections, a linear
For expected ranges of stack operation, it is desired to pick a few dynamic model is identified at each operating point, and scheduling
OPs whose dynamics can be used to span the the entire operating functions are identified that blend the linearized dynamics between
space. To select such points, a series of steady-state simulations operating points.
212 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217

Table 1
Selected OPs data.

OP num. Cell voltage (V) Fuel flow (mg s−1 )

(a) Input data

1 0.78 2.5
2 0.77 3.85
3 0.72 3.85
4 0.74 5.20
5 0.69 5.20
6 0.69 6.55

OP num. Cell current (A) H2 (mole fraction)

(b) Output data

Fig. 7. A postulated scheduling function fi ().
1 14.51 0.047
2 19.07 0.096
3 23.10 0.032
As an alternative to developing a parameterized basis expan-
4 24.96 0.102
5 30.31 0.038 sion of fi (), estimates of fi () can be found directly in terms of the
6 34.31 0.070 function values. Specifically, the scheduling parameter  is quan-
tized into a fixed number of “bins” with centers ¯ j (j = 1, 2, . . . , ),
with the objective being to estimate the values for fi (¯ j ) for all i,
3.1.1. Identification of linear models j as illustrated schematically in Fig. 7. In other words, the set of
At each OP in Table 1(a), a linear dynamic model is identified values fi (¯ j ) for all i, j are parameters to be identified. Note that 
that captures the dominant small-signal dynamics at that point. can be much larger than m. Once the values of fi (¯ j ) are found, the
This can be accomplished using linearization followed by model functions fi () can be represented by an appropriate interpolation.
reduction. However, the present approach uses the data-based Given a measured scheduling sequence  k , the quantized schedul-
model reduction method discussed in the companion paper [1]. The ing sequence ¯ k is established such that for all k, ¯ k = ¯ j for some j,
data-based identification is accomplished with a series of straight- and the Euclidean norm k − ¯ k  is minimized.
forward simulations, and is essentially independent of complexity Because the small-signal models Gi (z) are known for i = 1, 2, . . .,
in the underpinning physical model. Transient simulations in which m, the signals
the input variables are systematically perturbed provides the data
from which the reduced-order is obtained. The input variables are ıyki = Gi (z)(1 − z −1 )uk (2)
cell voltage, fuel mass flow rate, and air mass flow rate. The output can be calculated, where uk are the actuation inputs used in the
variables are cell current, hydrogen mole fraction in anode exhaust, simulations. The simulation outputs yk and quantized scheduling
and cathode-exhaust air temperature. The sampling time of the sequence ¯ k are also known. The objective is to select values of fi (¯ j )
perturbations is Ts = 0.125 s, which is also the sampling time of the such that ıyk ≈ ıŷk , where ıyk is the first difference of the measured
controller that is being designed. To take advantage of possible output
repeated modes between different OPs, the small-signal models are
assembled into a single model, as illustrated in Fig. 6. Thus, the out- ıyk = (1 − z −1 )yk , (3)
put of the identified system may be written as ıy = [ıy1 ıy2 · · · ıym ] and
where ıyi is the small-signal response of the system at OPi .

m

ıŷk = fi (¯ k )ıyki . (4)

3.2. Identification of scheduling functions
i=1

Using the results of the previous section, the small-signal mod- In other words, the objective is to find
els, Gi (z) (i = 1, 2, . . ., m), are available for m = 6 different OPs. In

N
other words, each Gi (z) captures the small-signal behavior for one argmin ıyk − ıŷk 2 , (5)
value of the scheduling parameter, . A set of scheduling functions fi (¯ j )
k=1
fi () (i = 1, 2, . . ., m) is needed to match the overall system response
to large-signal variations. As shown in Fig. 4, the functions fi () are so that the scheduling functions blend the small-signal models to
weighting factors that combine the outputs of the linearized mod- best match the observed output. However, without placing restric-
els. This means that fi () describes how closely model i describes tions on fi (¯ j ), the minimization problem is under-determined,
the model dynamics at scheduling point , and thus can be thought meaning that many different functions can make the objective
of as a correlation metric between the current operating condition function equal zero. Thus, constraints are required. The imposed
and operating point i. constraints are natural choices. One set of constraints come
from the definition of the scheduling functions: they must range
between 0 and 1 (0 ≤ fi (¯ j ) ≤ 1), and the sum of  the scheduling
m
functions over i for any ¯ j must always be unity ( i=1 fi (¯ j ) = 1).
Moreover, the value of each scheduling function at its correspond-
ing OP should be equal to unity, i.e. fi (¯ j = i ) = 1, where i is the
value of ¯ j that corresponds to the OP for system Gi (z). However, the
optimization problem is still under-determined. Including a regu-
larization term in the objective function solves the problem. This
term makes the identified functions sufficiently smooth. As intro-
duced by Hsu et al. [19], a dispersion function is used as a measure
of smoothness for the non-parametric identification of nonlinear
systems. The dispersion function is a smoothness measure for func-
Fig. 6. Combined linear model at multiple OPs. tions that are described point-wise, and is closely related to total
 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217 213

Fig. 8. Large-signal variation simulation.

variation. It can also be extended to multi-dimensional functions, 3.2.1. LPV-based nonlinear identification
¯ Given a point-wise
but for simplicity it is restricted here to scalar . Consider an example to illustrate the use of an LPV scheme for
description of a function in terms of points (¯ j , f (¯ j )), the dispersion the nonlinear system identification of the SOFC stack. Cell current is
of f is defined as chosen as the scheduling parameter. Using other output variables,
or combinations of output variables, as the scheduling variable is

−1
discussed in the following section. The stack OPs are specified by
 (f (¯ j )) = Lj2 , (6) two input variables, cell voltage and fuel flow rate. To transition
j=1 between different OPs, a transient simulation is designed in which
the output variables vary greatly within the acceptable operating
where Lj are the lengths of a linear interplant of (¯ j , f (¯ j ) (Fig. 7). The space (Fig. 5). Fig. 8a and b shows the temporal variation of the
dispersion is a quadratic function of the values of fi (¯ j ). Adding the input signals fuel flow rate and cell voltage. Fig. 8c and d shows
constraints and adding the smoothness measure to the objective the corresponding output responses for cell current and H2 in the
function, the optimization problem can be written as anode exhaust. To assist understanding how large-signal variations
move within the region defined by selected OPs, Fig. 9 illustrates

N

m
the changes in input and output spaces.
min ıyk − ıŷk 2 + ˇ  (fi (¯ j )) By solving the optimization problem as described by Eq. (7),
fi (¯ j )
k=1 i=1 six scheduling functions are identified. Fig. 10 illustrates the six
subject to 0 ≤ fi (¯ j ) ≤ 1, scheduling functions where cell current is used as the scheduling
(7)

m
parameter. A value of ˇ = 1 is used. As required by the constraints,
¯j
fi ( ) = 1,
i=1
fi (i ) = 1,

where ˇ is a user-defined parameter that can be varied to affect

the smoothness of the scheduling functions. Because the disper-
sion is a quadratic function of fi (¯ j ), Eq. (7) is a convex optimization
problem that can be easily solved using modern optimization pack-
ages as a second order cone problem. The results in this paper for
implementing the LPV algorithm were obtained using CVX [20,21]
toolbox in the computational software Matlab. By solving the opti-
mization problem, the values of the scheduling functions fi (¯ j ) are
calculated for all i, j that minimize a weighted sum of the fit error
and the dispersion, subject to the constraints.
By augmenting the linear state-space representation with the
past input and output (uk−1 and yk−1 ), the final result is a
parameter-varying state-space model of the form
      
uk 0 0 0 uk−1 I
xk+1 = − ˚ 0 xk +  uk , (8)
yk −D() C() I yk−1 D()

where

C() = [f1 () f2 () · · · fm ()]C, (9)

D() = [f1 () f2 () · · · fm ()]D, (10)

and (˚,  , C, D) are state-space matrices associated with the Fig. 9. Large-signal simulation. (a) Variation in input space over selected OPs. (b)
dynamics of the combined model G(z). Variation in output space over selected OPs.
214 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217

Fig. 11. Comparison between outputs from physically based high-order model and
the LPV-based low-order model over a wide range of operation. The results are
essentially indistinguishable.

the scheduling functions that are estimated when using hydrogen

in the anode exhaust as the scheduling parameter. It is also possible
to use more than one variable for scheduling. For instance, a combi-
Fig. 10. Identified scheduling functions using cell current as the scheduling variable.
nation of both cell current and exhaust hydrogen can be considered.

 values all satisfy 0 ≤ fi () ≤ 1, and at each value of cell

the function
current i fi () =1. Because the OPs are labeled in the order of cell
current (Table 1(b)), the functions have peaks in order from f1 () to
f6 () as the output current (i.e., ) increases. For example, because
OP2 has an output current of about 20 A, at  ≈ 20 A all the functions
are nearly zero except f2 ().
Within the LPV structure, the integrator is modified by adding a
simple error-correction term (predictor) as

ỹk = ỹk−1 + ıŷk + L(yk−1 − ỹk−1 ), (11)

where ỹk is the output of the LPV model, yk is the output of the actual
system (here the physical model) and ıŷk is calculated as in Eq.
(4). The objective is to reduce error accumulation in the integrator.
Since the model is intended to be used for real-time control, the
value of the system output yk−1 is available from actual sensors,
or sensor inferences. In Eq. (11), the prediction parameter L is a
free parameter, which can be chosen based on the length of the
experiment and the output variable. A value of 0.05 is chosen in
the results presented here.
The fidelity of the reduced-order models can be evaluated by
direct comparison with the detailed physical model. Fig. 11 shows
the results of both the low-order identified model and the full phys-
ical model, both being driven by the input transients shown in
Fig. 8a and b. To within the thickness of the lines, the two mod-
els produce essentially indistinguishable results. The LPV method,
based on smooth combinations of small-signal linear models, deliv-
ers an excellent representation of the nonlinear physical behavior
over wide ranges of operating conditions.

3.3. Selection of scheduling variable

Although cell current was used as the LPV scheduling parameter Fig. 12. Identified scheduling functions with H2 mole fraction in anode exhaust as
in the previous example, it is not the only choice. Fig. 12 illustrates the scheduling variable.
 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217 215

understanding into real-time process-control decisions. The SOFC

stack is a complex nonlinear system, with widely disparate ranges
of characteristic time scales. Moreover, there are multiple actua-
tion possibilities and a variety of possible sensors. It is a challenging
task to design and implement a control strategy that achieves opti-
mal performance through the coordination of multiple sensors and
actuators.
The MPC design is accomplished in two distinct parts: (1) esti-
mate the current system state from observations of past inputs u
and outputs y, and (2) establish future actuation trajectories u to
guide the system through a desired output trajectory.

4.1. State estimation

Assume that the process started at time k = 0, and the current

time is k = k0 . The first task of an MPC controller is to use recorded
sensor measurements yk (k = 0, 1, . . ., k0 ) to determine an estimate
for the current state of the plant, xk0 . Because actuation commands
are also recorded over time, the input–output information (uk ,
yk ) (k = 0, 1, . . ., k0 ) is available at time k0 . At any time step k, the
differences in actuation commands and output measurements can
be calculated as ıuk = uk − uk−1 and ıyk = yk − yk−1 , respectively. The
current state xk0 can be estimated by solving the following opti-
Fig. 13. Two-dimensional scheduling function as identified using cell current and mization problem:
exhaust H2 as the scheduling variables at OP1 . The function f1 ( 1 ,  2 ) ≈ 1 is highest
k 
in the left-lower corner, which corresponds to the cell current and exhaust H2 for
0 
k0 −1
OP1 . 2 2
min ek  + wk  , (12)
xk
k=0 k=0
In this case, the scheduling functions fi ( 1 ,  2 ) are two-dimensional
functions which must be identified by solving a slightly-modified where
optimization problem, specifically, two-dimensional dispersion




ek = Re ıyk − C  xk + D  ıuk , (13)
functions are employed as described in [19]. Intuitively, one might 

e
expect that the extended approach would deliver a higher fidelity wk = Q xk+1 − ˚xk + ıuk , (14)
identification, and thus improve performance of the reduced
model. Fig. 13 is a two-dimensional contour representation of f1 ( 1 , and Qe and Re
are weighing matrices. The state-estimation pro-
 2 ) corresponding to OP1 (Table 1(b)), where both cell current cess can be interpreted as follows: (1) find the state sequence
and exhaust hydrogen are the scheduling variables. The schedul- x̂k (k = 0, 1, . . . , k0 ) that minimizes a weighted combination of the
ing function is essentially unity (i.e., f1 ( 1 ,  2 ) ≈ 1) in the left lower error ek between the (small-signal) measurements (sensors) and
corner, which corresponds to the values of cell current and exhaust the model outputs and the adjustment term wk , needed to match
H2 around OP1 . Performance at other OPs is similar. the dynamic updates, and (2) take the last element of the state
Based on the simulation results for the present problem, there sequence x̂k as the current state of the plant, x̂k0 . The error (ek , Eq.
is little, if any, benefit in using the two-dimensional scheduling. (13)) can be interpreted in the context of the reduced-order model
Table 2 shows the root-mean-square (RMS) error associated with (Eq. (8)). At each scheduling variable value , the sensors measure
the alternative scheduling approaches. It is interesting to note the difference in output between two successive steps as ıyk while
that scheduling based on current provides a slightly better result the model predicts the difference as C()xk + D()ıuk . The adjust-
than scheduling with exhaust hydrogen. Based on these results, it ment wk is the difference between the state xk+1 at the next step
appears that the two scheduling variables are correlated. Thus, for and model-predicted state at the next step, ˚xk +  ıuk .
SOFC stack control, cell current is preferable because it is much The relative weighting matrices Qe and Re can be chosen based
easier to measure in practice. upon engineering judgment, or they can be chosen based on a
stochastic model for measurement errors and disturbances. In
either case, state estimation is a quadratic optimization problem. If
4. Model predictive control
the system is observable, the minimizer is unique. In this example,
Qe and Re are scaled identity matrices as Re = 0.1Ip and Qe = In where
The need to maintain signal constraints during operation,
p is the number of outputs and n is the number of states. The size of
combined with the importance of unmeasured variables such as
optimization problem (Eq. (12)) grows significantly as k0 increases.
internal stack temperature or fuel utilization, indicate the need
Fortunately, the Kalman filter can implement the estimation pro-
for advanced control strategies. Because of these important oper-
cess recursively [22]. That is, a gain sequence Kk can be calculated
ating considerations, as well as the strong interaction between
such that the output of the system,
input variables, MPC is a natural choice for control implementa-
tion. MPC provides a means to incorporate quantitative physical x̂k+1 = ˚x̂k + uk + Kk [ıyk − (C()xk + D()ıuk )], (15)

produces the same sequence of estimates for x̂k .

Table 2
Comparison between different scheduling approaches.
4.2. Control
Scheduling variable RMS error (%)

Cell current alone 0.16504 The role of control is to choose future actuation that guides the
Exhaust H2 alone 0.17632 system according to a desired trajectory. Assume that a desired
Cell current and exhaust H2 0.16121
output trajectory can be specified as ykd for k = k0 , . . ., k0 + p. The
216 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217

control is chosen to best match the desired output trajectory over

a look-ahead horizon of p sampling intervals, while meeting fixed
constraints on input and output excursions. At every time k0 the
current state x̂k0 is estimated according to the foregoing optimiza-
tion problem (i.e., Eq. (12)) and the system output at time k0 is
measured. After choosing weighting matrices Rc and Qc and con-
straints bui and byi , the following optimization problem is solved:

k0 +p
min Rc (ykd − yk )1 + Q c uk 1
uk
k=k0
subject to xk+1 = (˚xk + uk ) (16)
yk = C()xk + D()ıuk + yk−1
bu1 < uk < bu2
by1 < yk < by2

and the first element uk0 of the optimal sequence uk (k = k0 , k0 + 1,

. . ., k0 + p) is applied as the plant input (control law) for the next
time interval. In the example that follows, the horizon is cho-
sen as p = 12. For a sampling interval of Ts = 0.125 s, the horizon is
1.5 s. Weighting matrices Qc and Rc are square diagonal matrices
with different weights for different input and output variables. In
this example, Q c = diag([ 2.5 × 107 103 ]) where 2.5 × 107 is the
weight on fuel flow rate and 103 is the weight on cell voltage, and
Rc = diag([ 30 200 ]) where 30 is the weight on cell current and
200 is the weight on exhaust H2 .

Fig. 14. (a) Comparison between the desired cell current and the cell current deliv-
4.3. Controller performance
ered by the MPC controller. At this scale, the controlled result is essentially within
a line thickness of the desired current. (b) An expanded section of the trajectory,
To evaluate performance, an MPC controller is used to control revealing some small differences between desired and achieved cell current.
the physical SOFC model through a specified transient trajectory of
desired output current, while also satisfying constraints. The con-
troller uses the low-order LPV-based model for state estimation variables. In actual operation, the manipulated voltage is the sum
and actuation sequences. Fig. 14a shows both the desired current of the voltages across each cell, but here the “average” cell voltage,
trajectory and the achieved cell current using the MPC controller. or total voltage divided by number of cells, is reported. The con-
At this scale, the two curves are nearly indistinguishable. Fig. 14b troller maintains the commanded cell voltage and fuel flow rate
is an expanded graph over a short time interval that reveals some within the defined bounds (shown as the dashed lines). Cathode-
small differences between the desired and achieved cell-current air inlet flow rate is also controlled. However, for the short time
histories. The desired current spans a significantly wide range, interval shown in Fig. 15, it is essentially unchanged. Fig. 15c and
over which the physical behavior is strongly nonlinear. Although d show the output variables. The anode-exhaust hydrogen mole
near step-changes in the desired current trajectory, some small fraction remains in the desired pre-specified bounds. Because of
overshoots are present, the MPC controller is delivering excellent long thermal time constants, the cathode-exhaust air temperature
performance. varies relatively little over the short 40 s time interval.
In addition to achieving the desired cell-current trajectory These results show that the MPC controller provides excellent
(Fig. 14), the controller is able to respect constraints on the input performance over a very demanding cell-current trajectory. More-
and output variables. Fig. 15a and b shows the controlled input over, the controller meets the load demand, while also respecting

Fig. 15. MPC controlled input commands and model-predicted responses. Pre-specified bounds are shown with dashed lines. (a) Controller-commanded cell voltage. (b)
Controller-commanded fuel flow rate. (c) Hydrogen mole fraction in the anode exhaust. (d) Cathode-air exhaust temperature.
 B.M. Sanandaji et al. / Journal of Power Sources 196 (2011) 208–217 217

constraints on input and output variables. Qualitatively, the con- put variables, results show that scheduling based on the cell current
troller is making some anticipated decisions. For example, when alone provides excellent results.
the load decreases at around 5 s, the cell voltage increases and the
fuel flow rate decreases. As the load demand increases, the cell volt- References
age decreases and fuel flow rate increases. However, it would be
difficult, if not impossible, to achieve the complex set of controller [1] A.M. Colclasure, B.M. Sanandaji, T.L. Vincent, R.J. Kee, Modeling and control
of tubular solid-oxide fuel cell systems. I. Physical models and linear model
commands with less capable control strategies such as a PID regula- reduction. J. Power Sources 196 (2010) 196–207.
tor. Moreover, controllers such as PID are not capable of respecting [2] R.J. Kee, H. Zhu, D.G. Goodwin, Proc. Combust. Inst. 30 (2) (2005) 2379–2404.
input and output constraints. [3] R.J. Kee, H. Zhu, A.M. Sukeshini, G.S. Jackson, Combust. Sci. Technol. 180 (6)
(2008) 1207–1244.
[4] T.L. Vincent, B.M. Sanandaji, A.M. Colclasure, H. Zhu, R.J. Kee, ECS Trans. 25
5. Summary and conclusions (2009) 1175–1184.
[5] D.J. Hall, R.G. Colclaser, IEEE Trans. Energy Convers. 14 (3) (1999) 749–753.
[6] J. Padulles, G.W. Ault, J.R. McDonald, J. Power Sources 86 (1–2) (2000) 495–500.
A systematic approach is used to incorporate physical knowl- [7] J.T. Pukrushpan, A.G. Stefanopoulou, H. Peng, Control of Fuel Cell Power Sys-
edge into the control of an SOFC stack. The process begins with tems: Principles, Modeling, Analysis and Feedback Design, Springer Verlag,
a transient, high-fidelity, physical model. Such models, are too 2004.
[8] J. Yang, X. Li, H.G. Mou, L. Jian, J. Power Sources 188 (2) (2009) 475–482.
computationally expensive to be considered for direct incorpora-
[9] J. Arriagada, P. Olausson, A. Selimovic, J. Power Sources 112 (1) (2002) 54–60.
tion into real-time control. Reduced-order locally linear models [10] H.B. Huo, X.J. Zhu, W.Q. Hu, H.Y. Tu, J. Li, J. Yang, J. Power Sources 185 (1) (2008)
are required. However, the reduced models must be predictive 338–344.
over wide operating ranges, and thus capture nonlinear behav- [11] X. Wang, B. Huang, T. Chen, J. Process Control 17 (2) (2007) 103–114.
[12] J. Hasikos, H. Sarimveis, P.L. Zervas, N.C. Markatos, J. Power Sources 193 (1)
iors. Model reduction here is accomplished using an LPV structure. (2009) 258–268.
Results show that by smoothly scheduling between operating con- [13] Q. Chen, L. Gao, R.A. Dougal, S. Quan, J. Power Sources 191 (2) (2009) 473–482.
ditions, nonlinear behaviors can be adequately represented by [14] Q. Li, W. Chen, Y. Wang, J. Jia, M. Han, J. Power Sources 194 (1) (2009) 338–348.
[15] M.A. Danzer, J. Wilhelm, H. Aschemann, E.P. Hofer, J. Power Sources 176 (2)
combinations of low-order locally linear models. At certain steady- (2008) 515–522.
state OPs that span the expected operating space, locally linear [16] J. Yang, X. Li, H.G. Mou, L. Jian, J. Power Sources 193 (2) (2009) 699–705.
models are identified by systematic perturbations of the high-order [17] F. Jurado, J. Power Sources 158 (1) (2006) 245–253.
[18] B.M. Sanandaji, T.L. Vincent, A.M. Colclasure, R.J. Kee, Proceedings of the ASME
physical model. An MPC controller is designed and implemented Dynamic Systems and Control Conference, 2009.
based on the reduced-order models. In addition to meeting load- [19] K. Hsu, K. Poolla, T.L. Vincent, IEEE Trans. Automatic Control 53 (11) (2008)
demand transients, the MPC controller can also be designed to 2497–2513.
[20] M. Grant, S. Boyd, CVX: Matlab software for disciplined convex programming,
satisfy constraints on actuation commands (e.g., cell voltage and version 1.21. http://cvxr.com/cvx, April 2010.
fuel flow rate) and on output responses (e.g., fuel utilization). To [21] J. Löfberg, Proceedings of the CACSD Conference, Taipei, Taiwan, 2004.
evaluate performance, the LPV-based MPC controller is used to con- [22] L. Ljung, System Identification: Theory for the User, 2nd edition, Prentice-Hall,
1999.
trol the physically based high-order model. Although scheduling
between low-order linear models can depend upon multiple out-